基于改进YOLOv8的轻量化荷叶病虫害检测模型

Lightweight model for detecting lotus leaf diseases and pests using improved YOLOv8

腐败病、叶斑病、病毒病、斜纹夜蛾等荷叶病虫害严重影响莲子的产量与品质。开展疫病叶片检测是防治荷叶病虫害的重要措施。该研究以提高对荷叶病虫害的检测精度、减少模型的计算规模、提升可部署性为目标,提出了一种基于改进YOLOv8的轻量化荷叶病虫害检测模型,同时,建立了一种考虑不同环境条件的荷叶病虫害数据集。首先,将YOLOv8颈部网络中的卷积模块(Conv)替换为GSConv,将C2f模块替换为VoV-GSCSP,形成了Slim-neck架构,使模型在保持较高识别准确性的基础上降低计算复杂度。同时,使用融合了EMA高效多尺度注意力机制的C2f_EMA模块替换主干网络中的C2f模块,提升模型对复杂环境中荷叶病虫害的特征提取能力。试验结果表明,建立的改进YOLOv8荷叶病虫害检测模型能够对荷叶病虫害进行有效检测,实现的平均精度均值(mean average precision,mAP)为89.3%,较基线模型提高了1.6个百分点;模型的参数量较基线模型降低了0.2 M,模型大小仅为5.6 MB。与其他主流检测模型相比,改进YOLOv8模型在检测精度、参数量和模型大小等方面表现出显著优势。将模型部署至Jetson Xavier NX和树莓派4B边缘计算设备上,模型实现的检测帧率分别为27和0.7帧/s,展现了良好的移动端部署前景。所提模型实现了对荷叶病虫害的精准识别,可为荷叶病虫害自动防治提供支撑。

Lotus leaf diseases and pests have seriously threatened the yield and quality of lotus seeds,such as rhizome rot,leaf spot,virus disease,and Spodoptera litura.The detection of diseased leaves has been the most important measure to prevent and control the lotus leaf diseases and pests.However,manual detection cannot fully meet the large-scale production in recent years,due to the subjective and inefficient experience.It is still lacking in professional knowledge,easily leading to missed or false detection.Therefore,automatic detection can be expected to improve the planting quality in the actual environment of lotus fields.In this study,a lightweight detection model was proposed for the lotus leaf disease and pest using improved YOLOv8.The detection accuracy was improved to reduce the calculation scale for the better deployability of the model.At the same time,a new dataset of lotus leaf disease and pests was established to consider the different environmental conditions.Firstly,the convolution module(Conv)in the YOLOv8 neck network was replaced with GSConv.The C2f module was replaced with the VoV-GSCSP to form a slim-neck architecture,in order to reduce the computational complexity of the model for the high recognition accuracy.The C2f_EMA module was integrated with the EMA efficient multi-scale attention mechanism.The C2f module was then replaced in the backbone network,in order to extract the features of lotus leaf pests and diseases in the complex environments.The experimental results show that the improved YOLOv8 detection model effectively detected the lotus leaf pests and diseases.The mean average precision(mAP)was achieved at 89.3%,which was 1.6 percentages higher than the baseline model;The number of parameters of the model was reduced by 0.2M,compared with the baseline model.The model size was only 5.6 MB.A comparison was performed on several mainstream one-stage target detection models:Faster R-CNN,SSD,YOLOv3,YOLOv4,YOLOv5,YOLOv7,YOLOv8,and YOLOv9.The results show that the improved YOLOv8 model shared significant advantages in the detection accuracy,number of parameters,and model size,compared with the rest mainstream models.Finally,the improved YOLOv8 model was deployed on the Jetson Xavier NX and Raspberry Pi 4B edge computing devices,where the frame rates were 27 and 0.7 frames/s,respectively.Compared with the YOLOv5 model,the frame rates increased by 8.9 and 0.3 frames/s,respectively.In terms of actual deployment performance,the improved YOLOv8 model performed better than the classic YOLOv5 model,indicating a better prospect for the mobile terminal deployment.The accurate identification can provide support to the automatic prevention and control of lotus leaf diseases and pests.